Normal Matrices Learning goal: to prove that normal matrices are exactly those which have a full set

of orthonormal eigenvectors. When we had real symmetric matrices, we could always diagonalize by using an orthonormal set of eigenvectors. The eigenvalues of symmetric matrices were always real. But what if we allow more general eigenvalues while still requiring the eigenvectors to be orthonormal? Question: for what kinds of matrices A can we factor A = UUH with unitary U? We answered this question one way when we allowed SVDs to consider the possibility that the two unitary matrices were different. Now we require them to be the same. Lets see how good we can do. The idea would be to take A and make it diagonal through this transformation, but we know that is usually impossible. But lets get a start. Every matrix A has at least one eigenvectorit may not have more than one (we have !0 1 0 $ seen examples like # 0 0 1 & which have a triple eigenvalue with only one associated # & # & "0 0 0 % eigenvector). Take this eigenvector u1 and put it as the first column of U1, and then complete U1 in any way we like to be unitary. What is U1HAU1? The first column of AU1 is 1u1. The rest of AU1 is junk. So the first column of U1H(AU1) is (1, 0, , 0), while the other columns are junk. " !1 junk % So with a unitary matrix U1 we can transform A into A1 which has the form $ where B1 ' #0 &

B1 is some n 1 n 1 matrix. Now here is the cool trick. Look at B1. It is a brand new matrix. It has an eigenvector. Call this eigenvector u2. Note that it is only n 1 dimensional, and has nothing to do with being an eigenvector of A! Now, as before, let u2 be the first column of a matrix U2, and fill the rest of its columns with whatever might be needed to make U2 and n 1 n 1 unitary matrix. "1 0 % Consider the matrix U 2 = $ . This matrix is unitaryits columns are clearly !' #0 U2 & orthonormal! Now we find U2HA1U2 = junk % " (1 junk 0 % " (1 junk % " 1 0 % " 1 0 % " (1 % $ "1 ' = = " ( junk % ' . We can $0 U ! H ' $ 0 B1 ' $ 0 U 2 ' $ 0 U 2 H ' $ 0 [ (2 u 2 junk ]' $ 0 $ 2 !& # ! &# # &# 2 &# & $ B2 ' ' #0 && # keep going by finding an eigenvector of B2 and so on, clearing out all the entries below the main diagonal of the matrix. We have proven Theorem: for any A, there is a sequence of unitary Ui such that UnHUn-1HU2HU1HAU1Un is upper triangular.

We can do slightly better than this. Note that the product of two unitary matrices is unitary: (U1U2)H(U1U2) = U2HU1HU1U2 and this clearly multiplies out to I. So we can combine the entire mess above to restate it as Theorem (Schurs lemma): for any A, we can find unitary U such that UHAU is upper triangular. Note that the upper triangular matrix displays its eigenvalues on its diagonal, and these must be the eigenvalues of A, since the operation A S-1AS for any S does not change the characteristic polynomial, hence does not change the eigenvalues. So now we know how well we can do with a general A and a unitary U. When is the resulting upper triangular matrix actually diagonal? Certainly this is true for symmetric matrices, and in fact this extends to Hermitian matrices. For if UHAU = T is upper triangular, than (UHAU)H = TH is lower triangular. But on the left we get UHAHU = UHAU = T. So T = TH and the only way an upper triangular matrix can also be lower triangular is if it is diagonal. In fact, we have diagonalized the Hermitian matrix A, so the diagonal matrix is the matrix of eigenvalueswhich is real because it equals its own Hermitian, which is in this case just its conjugate. (Alternate explanation: if x is an eigenvector of Hermitian A with eigenvalue , then xHAx = xHx = ||x||2. But we have already seen that the expression on the left is real because it equals its own conjugate. Thus must be real.) The triangular matrix is also diagonal for skew-Hermitian A. Only now TH = -T. This shows that it is diagonal, with purely imaginary entries on the diagonal. The eigenvalues of a skew-Hermitian (or real skew-symmetric) matrix are purely imaginary. Another time we get a diagonal T is when A is itself unitary. This time (UHAU)H is actually the inverse of (UHAU). Thus TH = T-1. But the inverse of an upper triangular matrix is upper triangular, so the lower triangular TH must also be upper triangularand thus diagonal. And the diagonal entriesthe eigenvaluesmust all be complex numbers of modulus one (since their conjugates are their reciprocals). What is the most general class of matrices that have diagonal T? Definition: a matrix is normal if AAH = AHA, that is, A commutes with its Hermitian. Theorem: a matrix is normal if and only if it possesses a complete set of orthonormal eigenvectors. Proof: if A possesses a complete set of orthonormal eigenvectors, then UHAU is diagonal. So A = UDUH. And AH = UDHUH. These certainly commute since AAH = UDUHUDHUH and UHU = I, giving UDDHUH = UDHDUH since the diagonal matrices obviously commute, and this last is AHA. Conversely, if AAH = AHA then TTH = THT. It is easy to check by direct multiplication that if a triangular matrix is also normal, it must be diagonal. For example, in computing the very first entry, T H T1,1 = t1,1t1,1 while TT H 1,1 = t1,1t1,1 + t1,2t1,2 + ! + t1,nt1,n and since all these extra terms are non-negative (theyre squares of moduli of the ts), they must all be zero. So the first row of T is zero to the left of the diagonal entry.

Reading: 10.1, 10.2 Problems: 10.2: 1, 2, 3, 4, 6, 7, 15, 16, 17, 18, 25 (with proof!), Use Schurs lemma to prove: If all eigenvalues of a matrix A have moduli less than one, then lim A n = 0. (Assuming convergence of the appropriate series) the eigenvalues of eA are the exponentials of the eigenvalues of Aeven if A is not diagonalizable. Extra credit: show the series converges for any matrix A.
n!"